blob: b296942ff7d5865cc90d107edf6823cc9bd674d8 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * fs/direct-io.c
3 *
4 * Copyright (C) 2002, Linus Torvalds.
5 *
6 * O_DIRECT
7 *
8 * 04Jul2002 akpm@zip.com.au
9 * Initial version
10 * 11Sep2002 janetinc@us.ibm.com
11 * added readv/writev support.
12 * 29Oct2002 akpm@zip.com.au
13 * rewrote bio_add_page() support.
14 * 30Oct2002 pbadari@us.ibm.com
15 * added support for non-aligned IO.
16 * 06Nov2002 pbadari@us.ibm.com
17 * added asynchronous IO support.
18 * 21Jul2003 nathans@sgi.com
19 * added IO completion notifier.
20 */
21
22#include <linux/kernel.h>
23#include <linux/module.h>
24#include <linux/types.h>
25#include <linux/fs.h>
26#include <linux/mm.h>
27#include <linux/slab.h>
28#include <linux/highmem.h>
29#include <linux/pagemap.h>
Andrew Morton98c4d572006-12-10 02:19:47 -080030#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070031#include <linux/bio.h>
32#include <linux/wait.h>
33#include <linux/err.h>
34#include <linux/blkdev.h>
35#include <linux/buffer_head.h>
36#include <linux/rwsem.h>
37#include <linux/uio.h>
38#include <asm/atomic.h>
39
40/*
41 * How many user pages to map in one call to get_user_pages(). This determines
42 * the size of a structure on the stack.
43 */
44#define DIO_PAGES 64
45
46/*
47 * This code generally works in units of "dio_blocks". A dio_block is
48 * somewhere between the hard sector size and the filesystem block size. it
49 * is determined on a per-invocation basis. When talking to the filesystem
50 * we need to convert dio_blocks to fs_blocks by scaling the dio_block quantity
51 * down by dio->blkfactor. Similarly, fs-blocksize quantities are converted
52 * to bio_block quantities by shifting left by blkfactor.
53 *
54 * If blkfactor is zero then the user's request was aligned to the filesystem's
55 * blocksize.
56 *
57 * lock_type is DIO_LOCKING for regular files on direct-IO-naive filesystems.
58 * This determines whether we need to do the fancy locking which prevents
59 * direct-IO from being able to read uninitialised disk blocks. If its zero
Jes Sorensen1b1dcc12006-01-09 15:59:24 -080060 * (blockdev) this locking is not done, and if it is DIO_OWN_LOCKING i_mutex is
Linus Torvalds1da177e2005-04-16 15:20:36 -070061 * not held for the entire direct write (taken briefly, initially, during a
62 * direct read though, but its never held for the duration of a direct-IO).
63 */
64
65struct dio {
66 /* BIO submission state */
67 struct bio *bio; /* bio under assembly */
68 struct inode *inode;
69 int rw;
Daniel McNeil29504ff2005-04-16 15:25:50 -070070 loff_t i_size; /* i_size when submitted */
Linus Torvalds1da177e2005-04-16 15:20:36 -070071 int lock_type; /* doesn't change */
72 unsigned blkbits; /* doesn't change */
73 unsigned blkfactor; /* When we're using an alignment which
74 is finer than the filesystem's soft
75 blocksize, this specifies how much
76 finer. blkfactor=2 means 1/4-block
77 alignment. Does not change */
78 unsigned start_zero_done; /* flag: sub-blocksize zeroing has
79 been performed at the start of a
80 write */
81 int pages_in_io; /* approximate total IO pages */
82 size_t size; /* total request size (doesn't change)*/
83 sector_t block_in_file; /* Current offset into the underlying
84 file in dio_block units. */
85 unsigned blocks_available; /* At block_in_file. changes */
86 sector_t final_block_in_request;/* doesn't change */
87 unsigned first_block_in_page; /* doesn't change, Used only once */
88 int boundary; /* prev block is at a boundary */
89 int reap_counter; /* rate limit reaping */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080090 get_block_t *get_block; /* block mapping function */
Linus Torvalds1da177e2005-04-16 15:20:36 -070091 dio_iodone_t *end_io; /* IO completion function */
92 sector_t final_block_in_bio; /* current final block in bio + 1 */
93 sector_t next_block_for_io; /* next block to be put under IO,
94 in dio_blocks units */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -080095 struct buffer_head map_bh; /* last get_block() result */
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
97 /*
98 * Deferred addition of a page to the dio. These variables are
99 * private to dio_send_cur_page(), submit_page_section() and
100 * dio_bio_add_page().
101 */
102 struct page *cur_page; /* The page */
103 unsigned cur_page_offset; /* Offset into it, in bytes */
104 unsigned cur_page_len; /* Nr of bytes at cur_page_offset */
105 sector_t cur_page_block; /* Where it starts */
106
107 /*
108 * Page fetching state. These variables belong to dio_refill_pages().
109 */
110 int curr_page; /* changes */
111 int total_pages; /* doesn't change */
112 unsigned long curr_user_address;/* changes */
113
114 /*
115 * Page queue. These variables belong to dio_refill_pages() and
116 * dio_get_page().
117 */
118 struct page *pages[DIO_PAGES]; /* page buffer */
119 unsigned head; /* next page to process */
120 unsigned tail; /* last valid page + 1 */
121 int page_errors; /* errno from get_user_pages() */
122
123 /* BIO completion state */
124 spinlock_t bio_lock; /* protects BIO fields below */
125 int bio_count; /* nr bios to be completed */
126 int bios_in_flight; /* nr bios in flight */
127 struct bio *bio_list; /* singly linked via bi_private */
128 struct task_struct *waiter; /* waiting task (NULL if none) */
129
130 /* AIO related stuff */
131 struct kiocb *iocb; /* kiocb */
132 int is_async; /* is IO async ? */
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800133 int io_error; /* IO error in completion path */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 ssize_t result; /* IO result */
135};
136
137/*
138 * How many pages are in the queue?
139 */
140static inline unsigned dio_pages_present(struct dio *dio)
141{
142 return dio->tail - dio->head;
143}
144
145/*
146 * Go grab and pin some userspace pages. Typically we'll get 64 at a time.
147 */
148static int dio_refill_pages(struct dio *dio)
149{
150 int ret;
151 int nr_pages;
152
153 nr_pages = min(dio->total_pages - dio->curr_page, DIO_PAGES);
154 down_read(&current->mm->mmap_sem);
155 ret = get_user_pages(
156 current, /* Task for fault acounting */
157 current->mm, /* whose pages? */
158 dio->curr_user_address, /* Where from? */
159 nr_pages, /* How many pages? */
160 dio->rw == READ, /* Write to memory? */
161 0, /* force (?) */
162 &dio->pages[0],
163 NULL); /* vmas */
164 up_read(&current->mm->mmap_sem);
165
Jens Axboeb31dc662006-06-13 08:26:10 +0200166 if (ret < 0 && dio->blocks_available && (dio->rw & WRITE)) {
Nick Pigginb5810032005-10-29 18:16:12 -0700167 struct page *page = ZERO_PAGE(dio->curr_user_address);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 /*
169 * A memory fault, but the filesystem has some outstanding
170 * mapped blocks. We need to use those blocks up to avoid
171 * leaking stale data in the file.
172 */
173 if (dio->page_errors == 0)
174 dio->page_errors = ret;
Nick Pigginb5810032005-10-29 18:16:12 -0700175 page_cache_get(page);
176 dio->pages[0] = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 dio->head = 0;
178 dio->tail = 1;
179 ret = 0;
180 goto out;
181 }
182
183 if (ret >= 0) {
184 dio->curr_user_address += ret * PAGE_SIZE;
185 dio->curr_page += ret;
186 dio->head = 0;
187 dio->tail = ret;
188 ret = 0;
189 }
190out:
191 return ret;
192}
193
194/*
195 * Get another userspace page. Returns an ERR_PTR on error. Pages are
196 * buffered inside the dio so that we can call get_user_pages() against a
197 * decent number of pages, less frequently. To provide nicer use of the
198 * L1 cache.
199 */
200static struct page *dio_get_page(struct dio *dio)
201{
202 if (dio_pages_present(dio) == 0) {
203 int ret;
204
205 ret = dio_refill_pages(dio);
206 if (ret)
207 return ERR_PTR(ret);
208 BUG_ON(dio_pages_present(dio) == 0);
209 }
210 return dio->pages[dio->head++];
211}
212
Zach Brown6d544bb2006-12-10 02:20:54 -0800213/**
214 * dio_complete() - called when all DIO BIO I/O has been completed
215 * @offset: the byte offset in the file of the completed operation
216 *
217 * This releases locks as dictated by the locking type, lets interested parties
218 * know that a DIO operation has completed, and calculates the resulting return
219 * code for the operation.
220 *
221 * It lets the filesystem know if it registered an interest earlier via
222 * get_block. Pass the private field of the map buffer_head so that
223 * filesystems can use it to hold additional state between get_block calls and
224 * dio_complete.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800226static int dio_complete(struct dio *dio, loff_t offset, int ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227{
Zach Brown6d544bb2006-12-10 02:20:54 -0800228 ssize_t transferred = 0;
229
230 if (dio->result) {
231 transferred = dio->result;
232
233 /* Check for short read case */
234 if ((dio->rw == READ) && ((offset + transferred) > dio->i_size))
235 transferred = dio->i_size - offset;
236 }
237
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238 if (dio->end_io && dio->result)
Zach Brown6d544bb2006-12-10 02:20:54 -0800239 dio->end_io(dio->iocb, offset, transferred,
240 dio->map_bh.b_private);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241 if (dio->lock_type == DIO_LOCKING)
Ingo Molnard8aa9052006-07-03 00:25:02 -0700242 /* lockdep: non-owner release */
243 up_read_non_owner(&dio->inode->i_alloc_sem);
Zach Brown6d544bb2006-12-10 02:20:54 -0800244
245 if (ret == 0)
246 ret = dio->page_errors;
247 if (ret == 0)
248 ret = dio->io_error;
249 if (ret == 0)
250 ret = transferred;
251
252 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253}
254
255/*
256 * Called when a BIO has been processed. If the count goes to zero then IO is
257 * complete and we can signal this to the AIO layer.
258 */
259static void finished_one_bio(struct dio *dio)
260{
261 unsigned long flags;
262
263 spin_lock_irqsave(&dio->bio_lock, flags);
264 if (dio->bio_count == 1) {
265 if (dio->is_async) {
Zach Brown6d544bb2006-12-10 02:20:54 -0800266 int ret;
Daniel McNeil29504ff2005-04-16 15:25:50 -0700267
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 /*
269 * Last reference to the dio is going away.
270 * Drop spinlock and complete the DIO.
271 */
272 spin_unlock_irqrestore(&dio->bio_lock, flags);
Daniel McNeil29504ff2005-04-16 15:25:50 -0700273
Zach Brown6d544bb2006-12-10 02:20:54 -0800274 ret = dio_complete(dio, dio->iocb->ki_pos, 0);
Daniel McNeil29504ff2005-04-16 15:25:50 -0700275
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 /* Complete AIO later if falling back to buffered i/o */
277 if (dio->result == dio->size ||
278 ((dio->rw == READ) && dio->result)) {
Zach Brown6d544bb2006-12-10 02:20:54 -0800279 aio_complete(dio->iocb, ret, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 kfree(dio);
281 return;
282 } else {
283 /*
284 * Falling back to buffered
285 */
286 spin_lock_irqsave(&dio->bio_lock, flags);
287 dio->bio_count--;
288 if (dio->waiter)
289 wake_up_process(dio->waiter);
290 spin_unlock_irqrestore(&dio->bio_lock, flags);
291 return;
292 }
293 }
294 }
295 dio->bio_count--;
296 spin_unlock_irqrestore(&dio->bio_lock, flags);
297}
298
299static int dio_bio_complete(struct dio *dio, struct bio *bio);
300/*
301 * Asynchronous IO callback.
302 */
303static int dio_bio_end_aio(struct bio *bio, unsigned int bytes_done, int error)
304{
305 struct dio *dio = bio->bi_private;
306
307 if (bio->bi_size)
308 return 1;
309
310 /* cleanup the bio */
311 dio_bio_complete(dio, bio);
312 return 0;
313}
314
315/*
316 * The BIO completion handler simply queues the BIO up for the process-context
317 * handler.
318 *
319 * During I/O bi_private points at the dio. After I/O, bi_private is used to
320 * implement a singly-linked list of completed BIOs, at dio->bio_list.
321 */
322static int dio_bio_end_io(struct bio *bio, unsigned int bytes_done, int error)
323{
324 struct dio *dio = bio->bi_private;
325 unsigned long flags;
326
327 if (bio->bi_size)
328 return 1;
329
330 spin_lock_irqsave(&dio->bio_lock, flags);
331 bio->bi_private = dio->bio_list;
332 dio->bio_list = bio;
333 dio->bios_in_flight--;
334 if (dio->waiter && dio->bios_in_flight == 0)
335 wake_up_process(dio->waiter);
336 spin_unlock_irqrestore(&dio->bio_lock, flags);
337 return 0;
338}
339
340static int
341dio_bio_alloc(struct dio *dio, struct block_device *bdev,
342 sector_t first_sector, int nr_vecs)
343{
344 struct bio *bio;
345
346 bio = bio_alloc(GFP_KERNEL, nr_vecs);
347 if (bio == NULL)
348 return -ENOMEM;
349
350 bio->bi_bdev = bdev;
351 bio->bi_sector = first_sector;
352 if (dio->is_async)
353 bio->bi_end_io = dio_bio_end_aio;
354 else
355 bio->bi_end_io = dio_bio_end_io;
356
357 dio->bio = bio;
358 return 0;
359}
360
361/*
362 * In the AIO read case we speculatively dirty the pages before starting IO.
363 * During IO completion, any of these pages which happen to have been written
364 * back will be redirtied by bio_check_pages_dirty().
365 */
366static void dio_bio_submit(struct dio *dio)
367{
368 struct bio *bio = dio->bio;
369 unsigned long flags;
370
371 bio->bi_private = dio;
372 spin_lock_irqsave(&dio->bio_lock, flags);
373 dio->bio_count++;
374 dio->bios_in_flight++;
375 spin_unlock_irqrestore(&dio->bio_lock, flags);
376 if (dio->is_async && dio->rw == READ)
377 bio_set_pages_dirty(bio);
378 submit_bio(dio->rw, bio);
379
380 dio->bio = NULL;
381 dio->boundary = 0;
382}
383
384/*
385 * Release any resources in case of a failure
386 */
387static void dio_cleanup(struct dio *dio)
388{
389 while (dio_pages_present(dio))
390 page_cache_release(dio_get_page(dio));
391}
392
393/*
394 * Wait for the next BIO to complete. Remove it and return it.
395 */
396static struct bio *dio_await_one(struct dio *dio)
397{
398 unsigned long flags;
399 struct bio *bio;
400
401 spin_lock_irqsave(&dio->bio_lock, flags);
402 while (dio->bio_list == NULL) {
403 set_current_state(TASK_UNINTERRUPTIBLE);
404 if (dio->bio_list == NULL) {
405 dio->waiter = current;
406 spin_unlock_irqrestore(&dio->bio_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407 io_schedule();
408 spin_lock_irqsave(&dio->bio_lock, flags);
409 dio->waiter = NULL;
410 }
411 set_current_state(TASK_RUNNING);
412 }
413 bio = dio->bio_list;
414 dio->bio_list = bio->bi_private;
415 spin_unlock_irqrestore(&dio->bio_lock, flags);
416 return bio;
417}
418
419/*
420 * Process one completed BIO. No locks are held.
421 */
422static int dio_bio_complete(struct dio *dio, struct bio *bio)
423{
424 const int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
425 struct bio_vec *bvec = bio->bi_io_vec;
426 int page_no;
427
428 if (!uptodate)
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800429 dio->io_error = -EIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700430
431 if (dio->is_async && dio->rw == READ) {
432 bio_check_pages_dirty(bio); /* transfers ownership */
433 } else {
434 for (page_no = 0; page_no < bio->bi_vcnt; page_no++) {
435 struct page *page = bvec[page_no].bv_page;
436
437 if (dio->rw == READ && !PageCompound(page))
438 set_page_dirty_lock(page);
439 page_cache_release(page);
440 }
441 bio_put(bio);
442 }
443 finished_one_bio(dio);
444 return uptodate ? 0 : -EIO;
445}
446
447/*
448 * Wait on and process all in-flight BIOs.
449 */
Zach Brown6d544bb2006-12-10 02:20:54 -0800450static void dio_await_completion(struct dio *dio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 /*
453 * The bio_lock is not held for the read of bio_count.
454 * This is ok since it is the dio_bio_complete() that changes
455 * bio_count.
456 */
457 while (dio->bio_count) {
458 struct bio *bio = dio_await_one(dio);
Zach Brown6d544bb2006-12-10 02:20:54 -0800459 /* io errors are propogated through dio->io_error */
460 dio_bio_complete(dio, bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462}
463
464/*
465 * A really large O_DIRECT read or write can generate a lot of BIOs. So
466 * to keep the memory consumption sane we periodically reap any completed BIOs
467 * during the BIO generation phase.
468 *
469 * This also helps to limit the peak amount of pinned userspace memory.
470 */
471static int dio_bio_reap(struct dio *dio)
472{
473 int ret = 0;
474
475 if (dio->reap_counter++ >= 64) {
476 while (dio->bio_list) {
477 unsigned long flags;
478 struct bio *bio;
479 int ret2;
480
481 spin_lock_irqsave(&dio->bio_lock, flags);
482 bio = dio->bio_list;
483 dio->bio_list = bio->bi_private;
484 spin_unlock_irqrestore(&dio->bio_lock, flags);
485 ret2 = dio_bio_complete(dio, bio);
486 if (ret == 0)
487 ret = ret2;
488 }
489 dio->reap_counter = 0;
490 }
491 return ret;
492}
493
494/*
495 * Call into the fs to map some more disk blocks. We record the current number
496 * of available blocks at dio->blocks_available. These are in units of the
497 * fs blocksize, (1 << inode->i_blkbits).
498 *
499 * The fs is allowed to map lots of blocks at once. If it wants to do that,
500 * it uses the passed inode-relative block number as the file offset, as usual.
501 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800502 * get_block() is passed the number of i_blkbits-sized blocks which direct_io
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 * has remaining to do. The fs should not map more than this number of blocks.
504 *
505 * If the fs has mapped a lot of blocks, it should populate bh->b_size to
506 * indicate how much contiguous disk space has been made available at
507 * bh->b_blocknr.
508 *
509 * If *any* of the mapped blocks are new, then the fs must set buffer_new().
510 * This isn't very efficient...
511 *
512 * In the case of filesystem holes: the fs may return an arbitrarily-large
513 * hole by returning an appropriate value in b_size and by clearing
514 * buffer_mapped(). However the direct-io code will only process holes one
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800515 * block at a time - it will repeatedly call get_block() as it walks the hole.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 */
517static int get_more_blocks(struct dio *dio)
518{
519 int ret;
520 struct buffer_head *map_bh = &dio->map_bh;
521 sector_t fs_startblk; /* Into file, in filesystem-sized blocks */
522 unsigned long fs_count; /* Number of filesystem-sized blocks */
523 unsigned long dio_count;/* Number of dio_block-sized blocks */
524 unsigned long blkmask;
525 int create;
526
527 /*
528 * If there was a memory error and we've overwritten all the
529 * mapped blocks then we can now return that memory error
530 */
531 ret = dio->page_errors;
532 if (ret == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 BUG_ON(dio->block_in_file >= dio->final_block_in_request);
534 fs_startblk = dio->block_in_file >> dio->blkfactor;
535 dio_count = dio->final_block_in_request - dio->block_in_file;
536 fs_count = dio_count >> dio->blkfactor;
537 blkmask = (1 << dio->blkfactor) - 1;
538 if (dio_count & blkmask)
539 fs_count++;
540
Nathan Scott3c674e72006-03-29 09:26:15 +1000541 map_bh->b_state = 0;
542 map_bh->b_size = fs_count << dio->inode->i_blkbits;
543
Jens Axboeb31dc662006-06-13 08:26:10 +0200544 create = dio->rw & WRITE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545 if (dio->lock_type == DIO_LOCKING) {
546 if (dio->block_in_file < (i_size_read(dio->inode) >>
547 dio->blkbits))
548 create = 0;
549 } else if (dio->lock_type == DIO_NO_LOCKING) {
550 create = 0;
551 }
Nathan Scott3c674e72006-03-29 09:26:15 +1000552
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553 /*
554 * For writes inside i_size we forbid block creations: only
555 * overwrites are permitted. We fall back to buffered writes
556 * at a higher level for inside-i_size block-instantiating
557 * writes.
558 */
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800559 ret = (*dio->get_block)(dio->inode, fs_startblk,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560 map_bh, create);
561 }
562 return ret;
563}
564
565/*
566 * There is no bio. Make one now.
567 */
568static int dio_new_bio(struct dio *dio, sector_t start_sector)
569{
570 sector_t sector;
571 int ret, nr_pages;
572
573 ret = dio_bio_reap(dio);
574 if (ret)
575 goto out;
576 sector = start_sector << (dio->blkbits - 9);
577 nr_pages = min(dio->pages_in_io, bio_get_nr_vecs(dio->map_bh.b_bdev));
578 BUG_ON(nr_pages <= 0);
579 ret = dio_bio_alloc(dio, dio->map_bh.b_bdev, sector, nr_pages);
580 dio->boundary = 0;
581out:
582 return ret;
583}
584
585/*
586 * Attempt to put the current chunk of 'cur_page' into the current BIO. If
587 * that was successful then update final_block_in_bio and take a ref against
588 * the just-added page.
589 *
590 * Return zero on success. Non-zero means the caller needs to start a new BIO.
591 */
592static int dio_bio_add_page(struct dio *dio)
593{
594 int ret;
595
596 ret = bio_add_page(dio->bio, dio->cur_page,
597 dio->cur_page_len, dio->cur_page_offset);
598 if (ret == dio->cur_page_len) {
599 /*
600 * Decrement count only, if we are done with this page
601 */
602 if ((dio->cur_page_len + dio->cur_page_offset) == PAGE_SIZE)
603 dio->pages_in_io--;
604 page_cache_get(dio->cur_page);
605 dio->final_block_in_bio = dio->cur_page_block +
606 (dio->cur_page_len >> dio->blkbits);
607 ret = 0;
608 } else {
609 ret = 1;
610 }
611 return ret;
612}
613
614/*
615 * Put cur_page under IO. The section of cur_page which is described by
616 * cur_page_offset,cur_page_len is put into a BIO. The section of cur_page
617 * starts on-disk at cur_page_block.
618 *
619 * We take a ref against the page here (on behalf of its presence in the bio).
620 *
621 * The caller of this function is responsible for removing cur_page from the
622 * dio, and for dropping the refcount which came from that presence.
623 */
624static int dio_send_cur_page(struct dio *dio)
625{
626 int ret = 0;
627
628 if (dio->bio) {
629 /*
630 * See whether this new request is contiguous with the old
631 */
632 if (dio->final_block_in_bio != dio->cur_page_block)
633 dio_bio_submit(dio);
634 /*
635 * Submit now if the underlying fs is about to perform a
636 * metadata read
637 */
638 if (dio->boundary)
639 dio_bio_submit(dio);
640 }
641
642 if (dio->bio == NULL) {
643 ret = dio_new_bio(dio, dio->cur_page_block);
644 if (ret)
645 goto out;
646 }
647
648 if (dio_bio_add_page(dio) != 0) {
649 dio_bio_submit(dio);
650 ret = dio_new_bio(dio, dio->cur_page_block);
651 if (ret == 0) {
652 ret = dio_bio_add_page(dio);
653 BUG_ON(ret != 0);
654 }
655 }
656out:
657 return ret;
658}
659
660/*
661 * An autonomous function to put a chunk of a page under deferred IO.
662 *
663 * The caller doesn't actually know (or care) whether this piece of page is in
664 * a BIO, or is under IO or whatever. We just take care of all possible
665 * situations here. The separation between the logic of do_direct_IO() and
666 * that of submit_page_section() is important for clarity. Please don't break.
667 *
668 * The chunk of page starts on-disk at blocknr.
669 *
670 * We perform deferred IO, by recording the last-submitted page inside our
671 * private part of the dio structure. If possible, we just expand the IO
672 * across that page here.
673 *
674 * If that doesn't work out then we put the old page into the bio and add this
675 * page to the dio instead.
676 */
677static int
678submit_page_section(struct dio *dio, struct page *page,
679 unsigned offset, unsigned len, sector_t blocknr)
680{
681 int ret = 0;
682
Andrew Morton98c4d572006-12-10 02:19:47 -0800683 if (dio->rw & WRITE) {
684 /*
685 * Read accounting is performed in submit_bio()
686 */
687 task_io_account_write(len);
688 }
689
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 /*
691 * Can we just grow the current page's presence in the dio?
692 */
693 if ( (dio->cur_page == page) &&
694 (dio->cur_page_offset + dio->cur_page_len == offset) &&
695 (dio->cur_page_block +
696 (dio->cur_page_len >> dio->blkbits) == blocknr)) {
697 dio->cur_page_len += len;
698
699 /*
700 * If dio->boundary then we want to schedule the IO now to
701 * avoid metadata seeks.
702 */
703 if (dio->boundary) {
704 ret = dio_send_cur_page(dio);
705 page_cache_release(dio->cur_page);
706 dio->cur_page = NULL;
707 }
708 goto out;
709 }
710
711 /*
712 * If there's a deferred page already there then send it.
713 */
714 if (dio->cur_page) {
715 ret = dio_send_cur_page(dio);
716 page_cache_release(dio->cur_page);
717 dio->cur_page = NULL;
718 if (ret)
719 goto out;
720 }
721
722 page_cache_get(page); /* It is in dio */
723 dio->cur_page = page;
724 dio->cur_page_offset = offset;
725 dio->cur_page_len = len;
726 dio->cur_page_block = blocknr;
727out:
728 return ret;
729}
730
731/*
732 * Clean any dirty buffers in the blockdev mapping which alias newly-created
733 * file blocks. Only called for S_ISREG files - blockdevs do not set
734 * buffer_new
735 */
736static void clean_blockdev_aliases(struct dio *dio)
737{
738 unsigned i;
739 unsigned nblocks;
740
741 nblocks = dio->map_bh.b_size >> dio->inode->i_blkbits;
742
743 for (i = 0; i < nblocks; i++) {
744 unmap_underlying_metadata(dio->map_bh.b_bdev,
745 dio->map_bh.b_blocknr + i);
746 }
747}
748
749/*
750 * If we are not writing the entire block and get_block() allocated
751 * the block for us, we need to fill-in the unused portion of the
752 * block with zeros. This happens only if user-buffer, fileoffset or
753 * io length is not filesystem block-size multiple.
754 *
755 * `end' is zero if we're doing the start of the IO, 1 at the end of the
756 * IO.
757 */
758static void dio_zero_block(struct dio *dio, int end)
759{
760 unsigned dio_blocks_per_fs_block;
761 unsigned this_chunk_blocks; /* In dio_blocks */
762 unsigned this_chunk_bytes;
763 struct page *page;
764
765 dio->start_zero_done = 1;
766 if (!dio->blkfactor || !buffer_new(&dio->map_bh))
767 return;
768
769 dio_blocks_per_fs_block = 1 << dio->blkfactor;
770 this_chunk_blocks = dio->block_in_file & (dio_blocks_per_fs_block - 1);
771
772 if (!this_chunk_blocks)
773 return;
774
775 /*
776 * We need to zero out part of an fs block. It is either at the
777 * beginning or the end of the fs block.
778 */
779 if (end)
780 this_chunk_blocks = dio_blocks_per_fs_block - this_chunk_blocks;
781
782 this_chunk_bytes = this_chunk_blocks << dio->blkbits;
783
784 page = ZERO_PAGE(dio->curr_user_address);
785 if (submit_page_section(dio, page, 0, this_chunk_bytes,
786 dio->next_block_for_io))
787 return;
788
789 dio->next_block_for_io += this_chunk_blocks;
790}
791
792/*
793 * Walk the user pages, and the file, mapping blocks to disk and generating
794 * a sequence of (page,offset,len,block) mappings. These mappings are injected
795 * into submit_page_section(), which takes care of the next stage of submission
796 *
797 * Direct IO against a blockdev is different from a file. Because we can
798 * happily perform page-sized but 512-byte aligned IOs. It is important that
799 * blockdev IO be able to have fine alignment and large sizes.
800 *
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800801 * So what we do is to permit the ->get_block function to populate bh.b_size
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802 * with the size of IO which is permitted at this offset and this i_blkbits.
803 *
804 * For best results, the blockdev should be set up with 512-byte i_blkbits and
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800805 * it should set b_size to PAGE_SIZE or more inside get_block(). This gives
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806 * fine alignment but still allows this function to work in PAGE_SIZE units.
807 */
808static int do_direct_IO(struct dio *dio)
809{
810 const unsigned blkbits = dio->blkbits;
811 const unsigned blocks_per_page = PAGE_SIZE >> blkbits;
812 struct page *page;
813 unsigned block_in_page;
814 struct buffer_head *map_bh = &dio->map_bh;
815 int ret = 0;
816
817 /* The I/O can start at any block offset within the first page */
818 block_in_page = dio->first_block_in_page;
819
820 while (dio->block_in_file < dio->final_block_in_request) {
821 page = dio_get_page(dio);
822 if (IS_ERR(page)) {
823 ret = PTR_ERR(page);
824 goto out;
825 }
826
827 while (block_in_page < blocks_per_page) {
828 unsigned offset_in_page = block_in_page << blkbits;
829 unsigned this_chunk_bytes; /* # of bytes mapped */
830 unsigned this_chunk_blocks; /* # of blocks */
831 unsigned u;
832
833 if (dio->blocks_available == 0) {
834 /*
835 * Need to go and map some more disk
836 */
837 unsigned long blkmask;
838 unsigned long dio_remainder;
839
840 ret = get_more_blocks(dio);
841 if (ret) {
842 page_cache_release(page);
843 goto out;
844 }
845 if (!buffer_mapped(map_bh))
846 goto do_holes;
847
848 dio->blocks_available =
849 map_bh->b_size >> dio->blkbits;
850 dio->next_block_for_io =
851 map_bh->b_blocknr << dio->blkfactor;
852 if (buffer_new(map_bh))
853 clean_blockdev_aliases(dio);
854
855 if (!dio->blkfactor)
856 goto do_holes;
857
858 blkmask = (1 << dio->blkfactor) - 1;
859 dio_remainder = (dio->block_in_file & blkmask);
860
861 /*
862 * If we are at the start of IO and that IO
863 * starts partway into a fs-block,
864 * dio_remainder will be non-zero. If the IO
865 * is a read then we can simply advance the IO
866 * cursor to the first block which is to be
867 * read. But if the IO is a write and the
868 * block was newly allocated we cannot do that;
869 * the start of the fs block must be zeroed out
870 * on-disk
871 */
872 if (!buffer_new(map_bh))
873 dio->next_block_for_io += dio_remainder;
874 dio->blocks_available -= dio_remainder;
875 }
876do_holes:
877 /* Handle holes */
878 if (!buffer_mapped(map_bh)) {
879 char *kaddr;
Jeff Moyer35dc8162006-02-03 03:04:27 -0800880 loff_t i_size_aligned;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881
882 /* AKPM: eargh, -ENOTBLK is a hack */
Jens Axboeb31dc662006-06-13 08:26:10 +0200883 if (dio->rw & WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 page_cache_release(page);
885 return -ENOTBLK;
886 }
887
Jeff Moyer35dc8162006-02-03 03:04:27 -0800888 /*
889 * Be sure to account for a partial block as the
890 * last block in the file
891 */
892 i_size_aligned = ALIGN(i_size_read(dio->inode),
893 1 << blkbits);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 if (dio->block_in_file >=
Jeff Moyer35dc8162006-02-03 03:04:27 -0800895 i_size_aligned >> blkbits) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 /* We hit eof */
897 page_cache_release(page);
898 goto out;
899 }
900 kaddr = kmap_atomic(page, KM_USER0);
901 memset(kaddr + (block_in_page << blkbits),
902 0, 1 << blkbits);
903 flush_dcache_page(page);
904 kunmap_atomic(kaddr, KM_USER0);
905 dio->block_in_file++;
906 block_in_page++;
907 goto next_block;
908 }
909
910 /*
911 * If we're performing IO which has an alignment which
912 * is finer than the underlying fs, go check to see if
913 * we must zero out the start of this block.
914 */
915 if (unlikely(dio->blkfactor && !dio->start_zero_done))
916 dio_zero_block(dio, 0);
917
918 /*
919 * Work out, in this_chunk_blocks, how much disk we
920 * can add to this page
921 */
922 this_chunk_blocks = dio->blocks_available;
923 u = (PAGE_SIZE - offset_in_page) >> blkbits;
924 if (this_chunk_blocks > u)
925 this_chunk_blocks = u;
926 u = dio->final_block_in_request - dio->block_in_file;
927 if (this_chunk_blocks > u)
928 this_chunk_blocks = u;
929 this_chunk_bytes = this_chunk_blocks << blkbits;
930 BUG_ON(this_chunk_bytes == 0);
931
932 dio->boundary = buffer_boundary(map_bh);
933 ret = submit_page_section(dio, page, offset_in_page,
934 this_chunk_bytes, dio->next_block_for_io);
935 if (ret) {
936 page_cache_release(page);
937 goto out;
938 }
939 dio->next_block_for_io += this_chunk_blocks;
940
941 dio->block_in_file += this_chunk_blocks;
942 block_in_page += this_chunk_blocks;
943 dio->blocks_available -= this_chunk_blocks;
944next_block:
Eric Sesterhennd4569d22006-04-01 01:10:13 +0200945 BUG_ON(dio->block_in_file > dio->final_block_in_request);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946 if (dio->block_in_file == dio->final_block_in_request)
947 break;
948 }
949
950 /* Drop the ref which was taken in get_user_pages() */
951 page_cache_release(page);
952 block_in_page = 0;
953 }
954out:
955 return ret;
956}
957
958/*
Jes Sorensen1b1dcc12006-01-09 15:59:24 -0800959 * Releases both i_mutex and i_alloc_sem
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 */
961static ssize_t
962direct_io_worker(int rw, struct kiocb *iocb, struct inode *inode,
963 const struct iovec *iov, loff_t offset, unsigned long nr_segs,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800964 unsigned blkbits, get_block_t get_block, dio_iodone_t end_io,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 struct dio *dio)
966{
967 unsigned long user_addr;
968 int seg;
969 ssize_t ret = 0;
970 ssize_t ret2;
971 size_t bytes;
972
973 dio->bio = NULL;
974 dio->inode = inode;
975 dio->rw = rw;
976 dio->blkbits = blkbits;
977 dio->blkfactor = inode->i_blkbits - blkbits;
978 dio->start_zero_done = 0;
979 dio->size = 0;
980 dio->block_in_file = offset >> blkbits;
981 dio->blocks_available = 0;
982 dio->cur_page = NULL;
983
984 dio->boundary = 0;
985 dio->reap_counter = 0;
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -0800986 dio->get_block = get_block;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987 dio->end_io = end_io;
988 dio->map_bh.b_private = NULL;
989 dio->final_block_in_bio = -1;
990 dio->next_block_for_io = -1;
991
992 dio->page_errors = 0;
Chen, Kenneth W174e27c2006-03-25 03:08:16 -0800993 dio->io_error = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 dio->result = 0;
995 dio->iocb = iocb;
Daniel McNeil29504ff2005-04-16 15:25:50 -0700996 dio->i_size = i_size_read(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997
998 /*
999 * BIO completion state.
1000 *
1001 * ->bio_count starts out at one, and we decrement it to zero after all
1002 * BIOs are submitted. This to avoid the situation where a really fast
1003 * (or synchronous) device could take the count to zero while we're
1004 * still submitting BIOs.
1005 */
1006 dio->bio_count = 1;
1007 dio->bios_in_flight = 0;
1008 spin_lock_init(&dio->bio_lock);
1009 dio->bio_list = NULL;
1010 dio->waiter = NULL;
1011
1012 /*
1013 * In case of non-aligned buffers, we may need 2 more
1014 * pages since we need to zero out first and last block.
1015 */
1016 if (unlikely(dio->blkfactor))
1017 dio->pages_in_io = 2;
1018 else
1019 dio->pages_in_io = 0;
1020
1021 for (seg = 0; seg < nr_segs; seg++) {
1022 user_addr = (unsigned long)iov[seg].iov_base;
1023 dio->pages_in_io +=
1024 ((user_addr+iov[seg].iov_len +PAGE_SIZE-1)/PAGE_SIZE
1025 - user_addr/PAGE_SIZE);
1026 }
1027
1028 for (seg = 0; seg < nr_segs; seg++) {
1029 user_addr = (unsigned long)iov[seg].iov_base;
1030 dio->size += bytes = iov[seg].iov_len;
1031
1032 /* Index into the first page of the first block */
1033 dio->first_block_in_page = (user_addr & ~PAGE_MASK) >> blkbits;
1034 dio->final_block_in_request = dio->block_in_file +
1035 (bytes >> blkbits);
1036 /* Page fetching state */
1037 dio->head = 0;
1038 dio->tail = 0;
1039 dio->curr_page = 0;
1040
1041 dio->total_pages = 0;
1042 if (user_addr & (PAGE_SIZE-1)) {
1043 dio->total_pages++;
1044 bytes -= PAGE_SIZE - (user_addr & (PAGE_SIZE - 1));
1045 }
1046 dio->total_pages += (bytes + PAGE_SIZE - 1) / PAGE_SIZE;
1047 dio->curr_user_address = user_addr;
1048
1049 ret = do_direct_IO(dio);
1050
1051 dio->result += iov[seg].iov_len -
1052 ((dio->final_block_in_request - dio->block_in_file) <<
1053 blkbits);
1054
1055 if (ret) {
1056 dio_cleanup(dio);
1057 break;
1058 }
1059 } /* end iovec loop */
1060
Jens Axboeb31dc662006-06-13 08:26:10 +02001061 if (ret == -ENOTBLK && (rw & WRITE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 /*
1063 * The remaining part of the request will be
1064 * be handled by buffered I/O when we return
1065 */
1066 ret = 0;
1067 }
1068 /*
1069 * There may be some unwritten disk at the end of a part-written
1070 * fs-block-sized block. Go zero that now.
1071 */
1072 dio_zero_block(dio, 1);
1073
1074 if (dio->cur_page) {
1075 ret2 = dio_send_cur_page(dio);
1076 if (ret == 0)
1077 ret = ret2;
1078 page_cache_release(dio->cur_page);
1079 dio->cur_page = NULL;
1080 }
1081 if (dio->bio)
1082 dio_bio_submit(dio);
1083
Zach Brown17a7b1d2006-12-10 02:20:56 -08001084 /* All IO is now issued, send it on its way */
1085 blk_run_address_space(inode->i_mapping);
1086
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 /*
1088 * It is possible that, we return short IO due to end of file.
1089 * In that case, we need to release all the pages we got hold on.
1090 */
1091 dio_cleanup(dio);
1092
1093 /*
1094 * All block lookups have been performed. For READ requests
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001095 * we can let i_mutex go now that its achieved its purpose
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 * of protecting us from looking up uninitialized blocks.
1097 */
1098 if ((rw == READ) && (dio->lock_type == DIO_LOCKING))
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001099 mutex_unlock(&dio->inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100
1101 /*
1102 * OK, all BIOs are submitted, so we can decrement bio_count to truly
1103 * reflect the number of to-be-processed BIOs.
1104 */
1105 if (dio->is_async) {
1106 int should_wait = 0;
1107
Jens Axboeb31dc662006-06-13 08:26:10 +02001108 if (dio->result < dio->size && (rw & WRITE)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 dio->waiter = current;
1110 should_wait = 1;
1111 }
1112 if (ret == 0)
1113 ret = dio->result;
1114 finished_one_bio(dio); /* This can free the dio */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 if (should_wait) {
1116 unsigned long flags;
1117 /*
1118 * Wait for already issued I/O to drain out and
1119 * release its references to user-space pages
1120 * before returning to fallback on buffered I/O
1121 */
1122
1123 spin_lock_irqsave(&dio->bio_lock, flags);
1124 set_current_state(TASK_UNINTERRUPTIBLE);
1125 while (dio->bio_count) {
1126 spin_unlock_irqrestore(&dio->bio_lock, flags);
1127 io_schedule();
1128 spin_lock_irqsave(&dio->bio_lock, flags);
1129 set_current_state(TASK_UNINTERRUPTIBLE);
1130 }
1131 spin_unlock_irqrestore(&dio->bio_lock, flags);
1132 set_current_state(TASK_RUNNING);
1133 kfree(dio);
1134 }
1135 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136 finished_one_bio(dio);
Zach Brown6d544bb2006-12-10 02:20:54 -08001137 dio_await_completion(dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138
Zach Brown6d544bb2006-12-10 02:20:54 -08001139 ret = dio_complete(dio, offset, ret);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140
1141 /* We could have also come here on an AIO file extend */
Jens Axboeb31dc662006-06-13 08:26:10 +02001142 if (!is_sync_kiocb(iocb) && (rw & WRITE) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 ret >= 0 && dio->result == dio->size)
1144 /*
1145 * For AIO writes where we have completed the
1146 * i/o, we have to mark the the aio complete.
1147 */
1148 aio_complete(iocb, ret, 0);
1149 kfree(dio);
1150 }
1151 return ret;
1152}
1153
1154/*
1155 * This is a library function for use by filesystem drivers.
1156 * The locking rules are governed by the dio_lock_type parameter.
1157 *
1158 * DIO_NO_LOCKING (no locking, for raw block device access)
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001159 * For writes, i_mutex is not held on entry; it is never taken.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160 *
1161 * DIO_LOCKING (simple locking for regular files)
Nathan Scott3fb962b2006-03-15 15:14:45 +11001162 * For writes we are called under i_mutex and return with i_mutex held, even
1163 * though it is internally dropped.
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001164 * For reads, i_mutex is not held on entry, but it is taken and dropped before
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165 * returning.
1166 *
1167 * DIO_OWN_LOCKING (filesystem provides synchronisation and handling of
1168 * uninitialised data, allowing parallel direct readers and writers)
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001169 * For writes we are called without i_mutex, return without it, never touch it.
Nathan Scott3fb962b2006-03-15 15:14:45 +11001170 * For reads we are called under i_mutex and return with i_mutex held, even
1171 * though it may be internally dropped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 *
1173 * Additional i_alloc_sem locking requirements described inline below.
1174 */
1175ssize_t
1176__blockdev_direct_IO(int rw, struct kiocb *iocb, struct inode *inode,
1177 struct block_device *bdev, const struct iovec *iov, loff_t offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001178 unsigned long nr_segs, get_block_t get_block, dio_iodone_t end_io,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 int dio_lock_type)
1180{
1181 int seg;
1182 size_t size;
1183 unsigned long addr;
1184 unsigned blkbits = inode->i_blkbits;
1185 unsigned bdev_blkbits = 0;
1186 unsigned blocksize_mask = (1 << blkbits) - 1;
1187 ssize_t retval = -EINVAL;
1188 loff_t end = offset;
1189 struct dio *dio;
Nathan Scott3fb962b2006-03-15 15:14:45 +11001190 int release_i_mutex = 0;
1191 int acquire_i_mutex = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192
1193 if (rw & WRITE)
Jens Axboeb31dc662006-06-13 08:26:10 +02001194 rw = WRITE_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195
1196 if (bdev)
1197 bdev_blkbits = blksize_bits(bdev_hardsect_size(bdev));
1198
1199 if (offset & blocksize_mask) {
1200 if (bdev)
1201 blkbits = bdev_blkbits;
1202 blocksize_mask = (1 << blkbits) - 1;
1203 if (offset & blocksize_mask)
1204 goto out;
1205 }
1206
1207 /* Check the memory alignment. Blocks cannot straddle pages */
1208 for (seg = 0; seg < nr_segs; seg++) {
1209 addr = (unsigned long)iov[seg].iov_base;
1210 size = iov[seg].iov_len;
1211 end += size;
1212 if ((addr & blocksize_mask) || (size & blocksize_mask)) {
1213 if (bdev)
1214 blkbits = bdev_blkbits;
1215 blocksize_mask = (1 << blkbits) - 1;
1216 if ((addr & blocksize_mask) || (size & blocksize_mask))
1217 goto out;
1218 }
1219 }
1220
1221 dio = kmalloc(sizeof(*dio), GFP_KERNEL);
1222 retval = -ENOMEM;
1223 if (!dio)
1224 goto out;
1225
1226 /*
1227 * For block device access DIO_NO_LOCKING is used,
1228 * neither readers nor writers do any locking at all
1229 * For regular files using DIO_LOCKING,
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001230 * readers need to grab i_mutex and i_alloc_sem
1231 * writers need to grab i_alloc_sem only (i_mutex is already held)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 * For regular files using DIO_OWN_LOCKING,
1233 * neither readers nor writers take any locks here
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 */
1235 dio->lock_type = dio_lock_type;
1236 if (dio_lock_type != DIO_NO_LOCKING) {
1237 /* watch out for a 0 len io from a tricksy fs */
1238 if (rw == READ && end > offset) {
1239 struct address_space *mapping;
1240
1241 mapping = iocb->ki_filp->f_mapping;
1242 if (dio_lock_type != DIO_OWN_LOCKING) {
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001243 mutex_lock(&inode->i_mutex);
Nathan Scott3fb962b2006-03-15 15:14:45 +11001244 release_i_mutex = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 }
1246
1247 retval = filemap_write_and_wait_range(mapping, offset,
1248 end - 1);
1249 if (retval) {
1250 kfree(dio);
1251 goto out;
1252 }
1253
1254 if (dio_lock_type == DIO_OWN_LOCKING) {
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001255 mutex_unlock(&inode->i_mutex);
Nathan Scott3fb962b2006-03-15 15:14:45 +11001256 acquire_i_mutex = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 }
1258 }
1259
1260 if (dio_lock_type == DIO_LOCKING)
Ingo Molnard8aa9052006-07-03 00:25:02 -07001261 /* lockdep: not the owner will release it */
1262 down_read_non_owner(&inode->i_alloc_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263 }
1264
1265 /*
1266 * For file extending writes updating i_size before data
1267 * writeouts complete can expose uninitialized blocks. So
1268 * even for AIO, we need to wait for i/o to complete before
1269 * returning in this case.
1270 */
Jens Axboeb31dc662006-06-13 08:26:10 +02001271 dio->is_async = !is_sync_kiocb(iocb) && !((rw & WRITE) &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272 (end > i_size_read(inode)));
1273
1274 retval = direct_io_worker(rw, iocb, inode, iov, offset,
Badari Pulavarty1d8fa7a2006-03-26 01:38:02 -08001275 nr_segs, blkbits, get_block, end_io, dio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276
1277 if (rw == READ && dio_lock_type == DIO_LOCKING)
Nathan Scott3fb962b2006-03-15 15:14:45 +11001278 release_i_mutex = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279
1280out:
Nathan Scott3fb962b2006-03-15 15:14:45 +11001281 if (release_i_mutex)
Jes Sorensen1b1dcc12006-01-09 15:59:24 -08001282 mutex_unlock(&inode->i_mutex);
Nathan Scott3fb962b2006-03-15 15:14:45 +11001283 else if (acquire_i_mutex)
1284 mutex_lock(&inode->i_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 return retval;
1286}
1287EXPORT_SYMBOL(__blockdev_direct_IO);